Post on 13-Jun-2018
A Comparison of Orthomosaic Software for Use with Ultra High
Resolution Imagery of a Wetland Environment
John W. Gross a
a Center for Geographic Information Science and Geography Department, Central Michigan University, Mt. Pleasant, MI 48859
Abstract
Unmanned aerial systems are becoming popular in numerous military, civil, and research
applications as a platform for the collection of ultra-high spatial resolution imagery in a cost-
effective, dynamic manner. The use of unmanned aerial systems, however, requires significant
CPU intensive pre-processing to convert hundreds of raw images into a single usable
representation of the earth’s surface. The goal of this study is to compare geometric accuracy,
visual quality, ease of use and cost of Photoscan Pro, Pix4D Pro Mapper and Microsoft Image
Composite Editor; three modern image stitching software packages which have the potential to
significantly decrease the amount of time and user input required for creating orthomosaics. 223
usable images with spatial resolutions of 1.26 cm were collected on 26 September 2014 using an
unmanned aerial system. Photoscan Pro, Pix4d Pro Mapper, and Image Composite Editor were
used to create orthomosaics from the individual images. All orthomosaics were georeferenced
using 54 ground control points collected during the field season. The geometric accuracy of the
orthomosaics was calculated using the root mean square error of 17 validation points. Statistical
analysis was used to examine the level of significance between the different software’s. Microsoft
Image Composite Editor had significantly fewer visual errors (Chi Square, p < .001) but, it had the
poorest geometric accuracy with an RMSE of 34.7 cm (Tukey-Kramer, p < 0.05). Photoscan
Professional had the most visual errors (Chi Square, p < 0.001), and an RMSE of 10.9 cm. Pix4D
Pro had the best geometric accuracy with an RMSE of 7.7 cm, however this was not found to be
statistically different from Photoscan (Tukey-Kramer, p > 0.05). These results suggest that there
is no single best choice in terms of image stitching software, and that software selection must be
made on the budget and geometric and visual quality requirements of the individual project.
1. Introduction
Unmanned aerial systems (UASs), have been used, in varying forms since the 1930’s,
primarily in military applications (http://www.science.smith.edu/cmet/safetycode/faq.html;
Degarmo, 2004). Since the early 2000’s, however, the possible capabilities of UAVs in civil
applications such as biological research, precision agriculture, and archeology have become well
documented (Knoth et al, 2013; Verhoeven 2011; Verhoeven 2012; Zhenkun et al, 2013). This
growth will likely continue as government regulations and safety practices adapt to meet
demands (Zweig et al., 2014). The primary use of UAS in civil applications is as a platform for
the collection of ultra-high resolution (sub-decimeter) imagery. The use of such miniaturized
platforms and consumer-grade sensor equipment, however, leads to the collection of hundreds of
individual images which typically lack, or it exists but is of questionable accuracy, much of the
ancillary information commonly used in traditional airphoto interpretation. This lack of accurate
information translates into the requirement for novel techniques to convert these hundreds of
images into a single usable representation of the earth’s surface.
Many UAS studies have utilized consumer grade available digital cameras (Laliberte et
al, 2008).The use of such cameras combined with the small UAS platform can potentially create
a number of problems. There can be significant variability in rotational and angular camera
position, degree of overlap, and illumination (Barazzetti et al, 2010). The low flight altitude of
the UAS in relation to changes in ground elevation can lead to significant perspective distortions
(Zhang et al, 2011). Also, the global positions systems (GPS) and inertial measurement units on
board the UAS are typically too inaccurate to be meaningful, leading to significant errors in the
external operating parameters (EO) of the imagery (Laliberte et al, 2007; Turner et al, 2012).
Such errors, combined with the large number of images, make conventional photogrammetric
techniques difficult, if not impossible.
Recent advancements in the fields of photogrammetry and computer vision have
produced a number of algorithms that have the potential to not only handle these issues, but to do
so in a highly automated fashion. Such algorithms are able to detect and match hundreds of
overlapping images, accurately estimate internal and external camera parameters, create point
cloud representations of the 3D surface, and combine everything into a single, seamless 3D
model or orthomosaic. One of the most notable sets of algorithms for this purpose is structure
from motion (SfM). SfM is of benefit because it does not require a priori knowledge of any
camera parameters or scene information (Choudhary, 2012; Westoby et al, 2012). SfM typically
utilizes scale invariant feature transform (SIFT) to locate important features known as keypoints.
SIFT locates keypoints in an image using a difference-of-Gaussian function (Lowe, 2004).
Keypoints are then matched in each image based on the minimization of Euclidian distance
(Lindeberg, 2012). These keypoints are then tracked from image to image enabling the accurate
estimation of both camera orientation as well as keypoint location (Westoby et al, 2012). These
keypoints become a dense cloud allowing for the construction of a 3D representation of the
scene.
Previous research has utilized a variety of homemade SfM scripts to varying degrees of
success. Laliberte et al. (2008) flew imagery over the Jornada Experimental Range in southern
New Mexico. They used Autopano pro, a SIFT based software, to generate keypoints. These
keypoints were included in a custom script known as PreSync which also incorporated a 1 m
digital orthoquad and a 10 m digital elevation model to adjust and improve existing EOs. The
original images as well as the updated EOs were then put into Leica Photogrammetric Suite to
generate the mosaic. They were able to obtain overall RMSE of 47.9 cm which in part was due to
a lack of differential correction on their GPS unit. Turner et al 2012 created an automated
technique using SIFT and SfM techniques to automate the mosaicking of imagery collected over
two sites of an Antarctic moss bed. They were able to achieve mean absolute total errors ranging
from .103 m to 1.247 m.
Since 2010 a number of commercial SfM software packages have become available. Two
of the most popular include Photoscan Professional (Photoscan), and Pix4D Mapper Pro
(Pix4D). Both these software have been successfully used in current research (Vallet et al. 2011;
Kung et a.l 2011a; Kung et al. 2011b; Verhoeven et al.2011; Verhoeven et al. 2012; Woodget et
al. 2014).
The goal of this research was to compare software packages for use with imagery
acquired by UAS. Two commercially available Sfm software packages Photoscan and Pix4D
were compared as well as Microsoft Image Composite Editor (ICE) a freeware available through
Microsoft, which does not create a 3D point cloud, but rather stiches the images together
directly. This analysis focused on four main components: ease of use, visual quality, geometric
accuracy, and cost. Previous comparisons have reported that Photoscan Pro was the more
accurate software based primarily on digital surface modeling results and not a quantitative
analysis of image quality. (Sona et al. 2014; Turner et al. 2014),
2. Data
969 images were collected on 26 September 2014 at Braeburn Marsh Preserve near Ann
Arbor Michigan, USA (~42° 16’ N, ~84° 4’W). Imagery was collected using a Cannon EOS 6D
digital single lens reflex camera with a 50mm fixed focal length lens at a flight of 100m.The
camera was mounted on a Leptron Avenger airframe (Leptron, Golden, CO). All mission
parameters and flight limits were monitored and controlled via a ground station using piccolo
command center (Cloud Cap Technology, Hood River, OR). After the removal of low quality
images and turns in the flight lines (which reduce the accuracy of the final product), only 224
images were retained for analysis, covering 15.5 acres with a spatial resolution of 1.26 cm. Due
to the low accuracy of the camera GPS (25 m) no GPS EXIF data was retained for the analysis.
Ground control points (GCP) were established prior in the field season using a R8 GNSS
RTK rover and a TCS3 hand held unit (Trimble Navigation Limited, Sunnyvale, California) with
a recorded horizontal accuracy of +- 2 cm (95% confidence interval). Each point was marked
with a metal pole and colored foam for easy recognition in the imagery. A total of 70 points were
collected.
3. Methods
3.1 Overview
The 224 images, and 53 GCPs were used to create an image mosaic. The image mosaic
was created using all three software packages; Photoscan Pro, Pix4D, and ICE. The resultant
mosaics were then subjected to geometric and visual quality assessments and statistical analysis.
A basic workflow of the methodology can be seen in figure 1.
Figure 1: Basic workflow of methodology
3.2 Photoscan Pro
Created by the Russian based company Agisoft in 2010, Photoscan Pro is a 3D model and
image stitching software package. It utilizes an adapted form of the popular structure from
motion technology known as the scale invariant feature transformation proposed by Lowe
(2004). At its core this process uses feature points, which are simply geometrically similar and
distinct regions in an image e.g. building corners or the top of light posts. These points are then
tracked across multiple images creating a series of connections between photographs
(Verhoeven, 2011). In addition, this algorithm allows Photoscan Pro to automatically and
accurately estimate a large number of internal and external camera parameters which previously
had to be known and entered manually. Utilizing such algorithms, software packages such as
Photoscan Pro are capable of matching images at the subpixel level (Woodget et al., 2014).
The Photoscan Pro workflow can be broken down into four basic steps: image alignment,
dense point cloud formation, mesh creation, and texture creation. Each of these steps are run
independent of each other, and aside from the inclusion of ground control points they can be run
with little to no user input. It is also important to note that these stages are all independent and
can be saved separately for later use or revision.
There are a number of parameters that must be defined, often with multiple options for
each parameter. These parameters give the user control over a number of crucial factors that
affect the overall quality of the final output, such as: the maximum number of tie-points to
include in the cloud point, what type of surface the imagery consists of, and how to handle any
gaps in the final model. Trial and error in combination with a careful review of user’s manual
was used to parametrize each step (. A complete list of the parameters can be seen in table 1.
Table 1: Parameter settings for Photoscan
Photoscan
Step Parameter Setting
align photos accuracy high
pair
preselection disabled
point limit 40000
build dense pointcloud quality low
depth filtering mild
build mesh surface type height field
source data dense cloud
polygon count medium
interpolation enabled
build texture mapping mode adaptive orthophoto
blending mode mosaic
texture size 4096
3.3 Pix4D
Pix4D is alternative orthomosaic software created in 2011 by a Swiss company of the
same name. The Pix4D workflow consists of three steps: initial processing, point cloud
densification, and DSM and orthomosaic generation. The user defined properties which guide the
quality, accuracy, and format of the final output are all handled through a processing options
dialogue box which must be set up prior to any processing steps. The options in this box refined
into five sections: initial processing, point cloud, DSM orthomosaic, additional outputs, and
resources. A complete list of the parameters can be seen in table 2.
Before any processing is done or parameters are set, Pix4D highly recommends the
inclusion of any GCP data that may be available. This data is especially important since no GPS
data was included in the EXIF data for the images. GCP data also helps reduce shift problems
and other errors which may be present in the final model. Each of the GCPs was located in 5 of
the images to help ensure image quality.
Table 2: Parameter settings for Pix4D
Pix4D
Step Parameter Setting
initial processing processing aerial nadir
feature extraction 1
optimization externals and all internals
output none selected
point cloud image scale 1/4 (multiscale on)
point density optimal
minimum number of matches 3
point cloud filters none
DSM/orthomosaic noise filtering on
surface smoothing on
export type geotiff
resources resources all available
3.4 Microsoft ICE
ICE is an advanced panoramic image stitcher freeware produced by Microsoft typically
used to create detailed panoramas from numerous individual images. Unlike the other software
packages, ICE does not create a separate 3D point cloud and cannot generate 3D images. For
these reason, and because it is not designed to be used in scientific research, it does not create the
same number of outputs as Photoscan and Pix4D (i.e. no dense point clouds or
meshes).Therefore number of steps and parameters is relatively limited. In addition ICE
currently has no way to implicitly include GCPs in program. ICE is run in 4 sequential steps:
import, stitch, crop (optional) and export. A complete list of the parameters can be seen in table
3.
Table 3: Parameter settings for ICE
ICE
Step Parameter Setting
import panorama type simple
camera motion auto-detect (planar)
stitch roll 0 degrees
export scale 100
file format JPEG
quality 100/high
3.5 Geometric Accuracy Assessment
To quantify geometric accuracy, root mean square error (RMSE) was used. RMSE is the
comparison of real world (ground truth) information to estimated (image derived) measurements.
In the case of this research 17 GCPs had their real world location compared to their estimated
location in all orthomosaics. RMSE calculations were conducted using the following equation:
√∑ ((𝑋𝑖−𝑋𝑖,𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑)
2+(𝑌𝑖−𝑌𝑖,𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑)
2)𝑛
𝑖=1
𝑛 eq. (1)
where n is the total number of samples, Xi and Yi are the X and Y measurements of point
i’s location in the image, and X i,measured and Yi,measured are the position of GCP i as measured in
the field. To determine statistical significance a one way stacked ANOVA was conducted in
MINITAB (version 14) with α = 0.05. Although the data was not normally distributed ANOVA
was still conducted because ANOVA is not very sensitive to deviations from normality, and is
more powerful than the Kruskal-Wallis test (McDonald, 2014). Contingent on the results of the
ANOVA post hoc Tukey-Kramer test were also conducted for each pairwise comparison.
3.6 Visual Quality Assessment
To quantify the visual quality of the images, 100 points were created using the ‘create
random points’ tool in ArcGIS 10.2.2. A 2.5 m (5 m diameter) circular buffer was created around
each point. These locations were then assigned to one of four landcover classes based on visual
interpretation: reed canary grass (RCG), typha marsh (TM), wet-mesic prairie (WMP), or woody
vegetation (WV). A binary classifier (1 or 0) was used to denote the presence of image artifacts
and image blur. Chi Square tests for independence were calculated between image groups (i.e.
Pix4D vs ICE), and Fishers exact tests were used within image groups (Pix4D WV vs Pix4D
TM) with Bonferroni corrections applied as necessary (McDonald, 2014).
4. Results and Discussion
4.1 Geometric Accuracy
The results of this analysis are found in figure 2 and table 4. Overall Pix4D had the
highest level of geometric accuracy (RMSE = 7.7 cm) compared to Photoscan (RMSE = 10.9
cm) and ICE (RMSE = 34.7 cm). Pix4D and Photoscan were found to be statistically
significantly better than ICE (p < 0.05), however, they were not statistically different from each
other (p > 0.05) (table 4). This is readily apparent when the x and y errors are viewed
graphically, as ICE has much greater errors in both the x and y direction (figure 2). It should be
noted that a certain amount of the increased error of Photoscan relative to Pix4D may be due in
part to the increased presence of image artifacts which made it more difficult to accurately
determine the geometric centers of the GCPs.
The accuracies achieved by Photoscan Pro and Pix4d are comparable to values reported
in previous literature using UAS imagery. For example, a study conducted in 2010 by Laliberte
et al found RMSE values of 11.95 cm, 20.17 cm, and 16.69 cm for images collected over three
study sites in southwestern Idaho.
Table 2: Tukey test results for geometric accuracy assessment. ANOVA p value = 0.001 Bold values indicate significance
Lower Boundary Center Upper Boundary
ICE vs Photoscan -0.187 -0.108 -0.030
ICE vs Pix4D -0.193 -0.114 -0.036
Pix4D vs Photoscan -0.084 -0.006 0.072
Figure 2: Geometric error in the x and y direction for the 17 validation points
4.2 Visual Quality
The results of this analysis are shown in figures 3 & 4 and tables 5, 6 and 7. Photoscan
had significantly more image artifacts (p < 0.001) with 89% of the sites showing some degree of
artifacts (figure 3). Photoscan also had the highest proportion of image artifacts in each class,
with p values less than .001 in all comparisons (table 5). ICE did not produce any image artifacts
in any of the sites (figure 3). The only class comparison which was not found to be significant
was the comparison of wet-mesic prairie in ICE vs Pix4D (p = 0.306). The p value of typha
marsh between ICE and Pix4D (p = 0.019) was considered significant due to its proximity with
regards to the target value and the documented conservative nature of Bonferroni corrections
(McDonald, 2014).
Pix4D had the highest overall incidences of image blur with 38% of the total sites (figure
4). Pix4D was found to be overall statically significantly different from Photoscan (p < 0.001)
and ICE (p = 0.005) but, ICE was not found to be significantly different from Photoscan (p =
0.259) (table 6). Additionally, most of the within class comparisons were found to statistically
insignificant (table 6).
Woody vegetation had the highest percentages of both image blur and artifacts across all
three mosaics (figures 3 & 4). However, there was no statically significant difference found
between the classes with regards to the proportion of sites with artifacts in any of the mosaics.
Only the woody vegetation was found to be statistically significant from the other classes in
Photoscan and Pix4D with regards to image blur (table 7). Again, the p value of 0.017 was
considered significant due to the conservative nature of Bonferroni corrections. One potential
cause of the increased number of sites with image blur and artifacts found in woody vegetation is
the increased presence of windblown leaf movement occurring in the canopies between flight
lines. This relative shift in location between images compared to other more stable targets, may
increase the difficulty of matching those portions of the images.
Table 5: p values for each class in terms of image artifacts compared between the three mosaics. Value in parenthesis is the target p value for the column after Bonferroni corrections bold values are significant. * indicates potential false negative.
Class ICE vs Photoscan
vs Pix4D (p=0.050)
ICE vs Photoscan (p=0.017)
ICE vs Pix4D (p=0.017)
Photoscan vs Pix4D (p=0.017)
Overall < 0.001 < 0.001 < 0.001 < 0.001
Reed canary grass < 0.001 < 0.001 < 0.001 < 0.001
Wet-mesic prairie < 0.001 < 0.001 0.306 0.001
Typha marsh < 0.001 < 0.001 0.019* 0.001
Woodland Vegetation < 0.001 < 0.001 < 0.001 < 0.001 Table 6: p values for each class in terms of image blur compared between the three mosaics. Value in parenthesis is the target p value for the column after Bonferroni corrections bold values are significant. NA represent comparisons that was not calculated
because overall comparison was not significant or did not have enough data.
Class ICE vs Photoscan
vs Pix4D (p=0.050)
ICE vs Photoscan (p=0.017)
ICE vs Pix4D (p=0.017)
Photoscan vs Pix4D (p=0.017)
Overall < 0.001 0.259 0.005 < 0.001
Reed canary grass 0.040 0.040 0.585 0.011
Wet-mesic prairie 0.192 NA NA NA
Typha marsh 0.081 NA NA NA
Woodland Vegetation 0.001 0.397 0.001 0.007
Table 7: p values for each class comparison in terms of image blur. Value in parenthesis is the target p value for the column after Bonferroni corrections bold values are significant. * indicates potential false negative. NA represent data that was not
calculated because overall comparison was not significant or did not have enough data.
class ICE
(0.008) Pix4D
(0.008) Photoscan
(0.008)
overall 0.065 < 0.001 < 0.001
canary reed grass
NA 0.399 1.000 vs
wet-mesic prairie
canary reed grass
NA 0.550 1.000 vs
typha marsh
canary reed grass
NA < 0.001 0.001 vs
woody vegetation
wet-mesic prairie
NA 1.000 1.000 vs
typha marsh
wet-mesic prairie
NA 0.003 0.017* vs
woody vegetation
typha marsh
NA < 0.001 0.002 vs
woody vegetation
Figure3: percent of each class with image artifacts Figure 4: percent of each class with image blur
4.3 Cost
Although ICE is a freely available download from Microsoft, it does not have the same
features as the other software, namely 3D surface modeling. However, Installation is easy and
expandable without complicated licensing software. In terms of the commercial SfM software
packages, Photoscan is significantly cheaper, costing only $549 for an educational license.
Comparatively Pix4D costs $4,990 for a non-commercial license. A less expensive option is
available, however, it has stringent requirements on publication processes and requires yearly
update reports with regards to use. Pix4D does offer license rentals at monthly or yearly rates of
$350 and $8,700 respectively, however, no educational discounts are available for these rates. It
is also important to note that Pix4D includes a second license specifically designated to using the
rapid check functionality in the field to verify data accuracy.
4.4 Ease of Use
ICE’s user interface (UI) was the simplest of the three programs, due in part to the fact
that it does not generate a 3D point cloud. Many of its parameters consist of slid bars, which
0.0
0
0.0
0
0.0
0
0.0
0
0.0
0
29
.00
31
.43
9.0
9
19
.23
42
.86
89
.00
88
.57
81
.82
84
.62
96
.43
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
80.00
90.00
100.00P
ERC
ENT
WIT
H A
RTI
FAC
TS
CLASS
ICE Pix4D Photoscan
20
.00
22
.86
27
.27
3.8
5
28
.57
38
.00
28
.57
18
.18
19
.23
75
.00
14
.00
5.7
1
0.0
0 3.8
5
39
.29
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
80.00
90.00
100.00
PER
CEN
T W
ITH
BLU
R
CLASS
ICE Pix4D Photoscan
make adjustments simple. With only a single processing step, ICE offers the least amount of
control over the workflow. Additionally, it does not possess any in program GCP placement UI,
meaning that third party software such as ENVI or ArcGIS is required to georeference the
resulting mosaic.
Overall the UI of Photoscan was clean and user friendly. Images could be easily added,
disabled, and removed at any point in the workflow. Each step could be run and saved
independently making it much easier to adjust settings of individual steps to achieve the best
results. Multiple GCPs could be placed in one image at a time, and have their location predicted,
significantly reducing time and effort when placing GCPs, which is the most user intensive step
of the process. Photoscan also offers an easy to use batch processing mode, which allows users to
chain together any functionality, set parameters, and run them as one process.
Pix4D offered the most complicated UI. Included in its UI is a basemap image allowing
for an in program comparison of GCPs and mosaics to the real world prior to export.
Unfortunately, this can, at times, make viewing the finished mosaic in program slightly
cumbersome. In contrast to Photoscan’s modular setup, Pix4D has all its parameters set up in
advance using a single options menu. The initial processing, point cloud densification, and
DSM/Orthomosaic creation can then be run independently or all handled in a single execution.
Pix4D highly recommends the inclusion of GCP data not only to georeference the final outputs,
but also to help avoid a number of common errors, such as inverted mosaics. Pix4D’s UI for
GCP placement, however, uses small sub windows within the layout and requires significant
time and energy to implement (approx. 8 hours for 52 GCPs with 5 images each), compared to
Photoscan (approx. 4 hours).
5. Conclusion
With the use of UAS increasing both in the academic and civil sectors it is important to
find quick, accurate methods for converting significant numbers of images into a single usable
orthomosaic. This research has presented three possible alternative software packages which
offer a highly automated approach, Photoscan Pro, Pix4D, and Microsoft Image Composite
Editor. A summary of the findings can be seen in table 8. Although, Photoscan Pro and
Microsoft Image Composite Editor were easier to use than Pix4D, Photoscan proved to be least
effective in terms of visual accuracy, and ICE had the worst geometric accuracy. In addition all
three software packages struggled with tree canopies where windblown leaf movement led to
increased numbers of image artifacts or blur. Therefore, this research suggests that there is no
single best option for optimizing all criteria, and a software selection should reflect the
cost/benefit trade off between cost, ease of use, geometric accuracy, and visual quality for the
individual application.
Table8: Overall comparison between the software. Numbers represent the software order for that category with 1 being the highest. Duplicate values were awarded in cases where no statistical difference was observed
6. Acknowledgements and Funding
This research was funded by the Center for Geographic Information Science at Central Michigan
University. Without the help of Dr. Benjamin Heumann, Rachel Hackett, Brian Stark, and Sam Lipscomb
with regards to data collection this project would not have been possible. In addition, a special thanks is
given to Braeburn Nature Preserve for allowing us frequent access to the site.
7. References
Barazzetti, L., F. Remondino, M. Scaioni, and R. Brumana. 2010. Fully automatic UAV image-
based sensor orientation.in Proceedings of the 2010 Canadian Geomatics Conference and
Symposium of Commission I.
Choudhary, S., and P. Narayanan. 2012. Visibility probability structure from sfm datasets and
applications. Pages 130-143 Computer Vision–ECCV 2012. Springer.
DeGarmo, M. T. 2004. Issues concerning integration of unmanned aerial vehicles in civil
airspace. The MITRE Corporation Center for Advanced Aviation System Development.
Knoth, C., B. Klein, T. Prinz, and T. Kleinebecker. 2013. Unmanned aerial vehicles as
innovative remote sensing platforms for high‐resolution infrared imagery to support
restoration monitoring in cut‐over bogs. Applied Vegetation Science 16:509-517.
Küng, O., C. Strecha, A. Beyeler, J.-C. Zufferey, D. Floreano, P. Fua, and F. Gervaix. 2011a.
The accuracy of automatic photogrammetric techniques on ultra-light UAV imagery.in
UAV-g 2011-Unmanned Aerial Vehicle in Geomatics.
Küng, O., C. Strecha, P. Fua, D. Gurdan, M. Achtelik, K.-M. Doth, and J. Stumpf. 2011b.
Simplified building models extraction from ultra-light uav imagery. ISPRS-International
Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
3822:217.
Laliberte, A. S., J. E. Herrick, A. Rango, and C. Winters. 2010. Acquisition, orthorectification,
and object-based classification of unmanned aerial vehicle (UAV) imagery for rangeland
monitoring. Photogrammetric Engineering & Remote Sensing 76:661-672.
Laliberte, A. S., A. Rango, and J. Herrick. 2007. Unmanned aerial vehicles for rangeland
mapping and monitoring: a comparison of two systems.in ASPRS Annual Conference
Proceedings.
Laliberte, A. S., C. Winters, and A. Rango. 2008. A procedure for orthorectification of sub-
decimeter resolution imagery obtained with an unmanned aerial vehicle (UAV). Pages
08-047 in Proc. ASPRS Annual Conf.
Geometric accuracy
Visual qualityCost
Ease of use
ICE 2 1 1 2
Photoscan Pro 1 3 2 1
Pix4D 1 2 3 3
Lindeberg, T. 2012. Scale invariant feature transform. Scholarpedia 7:10491.
Lowe, D. G. 2004. Distinctive image features from scale-invariant keypoints. International
journal of computer vision 60:91-110.
McDonald, J. H. 2009. Handbook of biological statistics. Sparky House Publishing Baltimore,
MD.
Sona, G., L. Pinto, D. Pagliari, D. Passoni, and R. Gini. 2014. Experimental analysis of different
software packages for orientation and digital surface modelling from UAV images. Earth
Science Informatics 7:97-107.
Turner, D., A. Lucieer, and L. Wallace. 2014. Direct georeferencing of ultrahigh-resolution uav
imagery.
Turner, D., A. Lucieer, and C. Watson. 2012. An automated technique for generating
georectified mosaics from ultra-high resolution unmanned aerial vehicle (UAV) imagery,
based on structure from motion (SfM) point clouds. Remote Sensing 4:1392-1410.
Verhoeven, G. 2011. Taking computer vision aloft–archaeological three‐dimensional
reconstructions from aerial photographs with photoscan. Archaeological Prospection
18:67-73.
Verhoeven, G., M. Doneus, C. Briese, and F. Vermeulen. 2012. Mapping by matching: a
computer vision-based approach to fast and accurate georeferencing of archaeological
aerial photographs. Journal of Archaeological Science 39:2060-2070.
Westoby, M., J. Brasington, N. Glasser, M. Hambrey, and J. Reynolds. 2012. ‘Structure-from-
Motion’photogrammetry: A low-cost, effective tool for geoscience applications.
Geomorphology 179:300-314.
Woodget, A., P. Carbonneau, F. Visser, and I. Maddock. 2014. Quantifying submerged fluvial
topography using hyperspatial resolution UAS imagery and structure from motion
photogrammetry. Earth Surface Processes and Landforms.
Zhang, Y., J. Xiong, and L. Hao. 2011. Photogrammetric processing of low‐altitude images
acquired by unpiloted aerial vehicles. The Photogrammetric Record 26:190-211.
Zhenkun, T., F. Yingying, and L. Suhong. 2013. Rapid crops classification based on UAV low-
altitude remote sensing. Transactions of the Chinese Society of Agricultural Engineering
2013.